Tian Lan, Yiqian Yang, Qianghuai Jia, Li Zhu, Hui Jiang, Hang Zhu, Weihua Luo, Longyue Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Voice agents can construct meaningful conversations with older adults to offer various benefits, such as providing emotional companionship and assisting with memory recall. However, such conversations often follow the simple turn-taking pattern and lack interruption and backchannel of natural human conversation. Previous research has shown that this rigid turn-taking pattern lacks interactivity and initiative, limiting the flexible communication between older adults and voice agents. To address these issues and create a more natural conversational voice agent, we first conducted a formative study to identify common usage of interruption in the natural conversations of older adults. We then designed an LLM-powered Barge-in agent that supports interruption and backchannel. Our within-subject exploratory study showed that participants felt that conversations with Barge-in agents were more natural, engaging, and fluent than with the No barge-in agent. We further present design implications for creating more natural and human-like voice agents for older adults.

Accurate modeling and control of response length is essential for optimizing large language model (LLM) deployment, impacting computational efficiency, user experience, and system reliability.We develop a statistical framework based on extreme value theory, analyzing 14,301 GPT-4o responses across temperature settings and prompting strategies, with cross-validation on Qwen and DeepSeek architectures.Our analysis reveals that response lengths follow Weibull-type generalized extreme value (GEV) distributions, exhibiting heavier tails under stochastic generation conditions.The key contributions include:(1) a novel GEV-generalized Pareto (GPD) hybrid model that achieves superior tail fit (R 2 CDF = 0.9993 vs standalone GEV's 0.998) while preserving architectural generalizability;(2) quantitative characterization of prompt anchoring effects, showing reduced dispersion but increased outlier propensity under randomization; and (3) identification of temperaturedependent response patterns that remain consistent across architectures, where higher temperatures amplify length variability while maintaining the underlying extreme-value mechanisms.The proposed hybrid model's adaptive threshold selection enables precise verbosity control in production systems, regardless of the specific LLM architecture employed.These findings provide both theoretical insights into LLM generation patterns and practical tools for response length optimization.

Incorporating both RGB and depth images has proven effective for enhancing the performance of semantic segmentation. However, current RGB-D semantic segmentation methods tend to overlook the critical role of cross-modal difference information during fusion, leading to the undesired suppression of discriminative cues and a failure to achieve potent cross-modal complementary fusion. In this article, a novel RGB-D semantic segmentation approach that realizes the efficient utilization of multimodal information is proposed. To address the issue of the suppression of cross-modal difference information, we propose a dynamic frequency-spatial difference-aware fusion module adept at explicitly emphasizing cross-modal differences, capturing vital features in the frequency domain, and using them to aggregate spatial context information of multimodal features. We also present a novel soft-edge loss to meticulously handle complex scenes by supervising different regions respectively. In addition, a progressive calibration context module is designed to enhance global contextual information by capturing multiscale multimodal representations. Extensive experiments on two public RGB-D datasets demonstrate that the proposed DFNet achieves highly competitive performance compared to state-of-the-art methods, making it well-suited for assisting indoor robots.

In autonomous driving systems, multispectral semantic segmentation integrates thermal images with RGB images to mitigate reliability degradation under challenging illumination conditions. However, most RGB-T methods cannot maintain high accuracy without sacrificing efficiency, thereby hindering efficient scene understanding. This paper presents RADNet, an asymmetric dual-stream network optimized for real-time RGB-T semantic segmentation, which incorporates three important components. The FFD module analyzes thermal features from the perspective of frequency to enhance thermal representations while suppressing interference. Furthermore, the ISD strategy explicitly transfers multimodal knowledge under varying illumination, leveraging the strengths of each modality without adding inference overhead. In addition, the FR module combines standard convolutions and central difference convolutions via structural re-parameterization to enhance detailed information. RADNet shows noticeable efficiency with 198.49 FPS on an RTX 3090 GPU while maintaining 13.21M parameters. Experimental results across three multispectral benchmarks covering diverse scenarios and illumination variations validate its competitive accuracy. Our network strikes a promising balance between segmentation performance and computational efficiency, making it a promising solution for further exploration in autonomous driving systems.

Current mainstream vision-language (VL) tracking framework consists of three parts,i.e., a visual feature extractor, a language feature extractor, and a fusion model. To pursue better performance, a natural modus operandi for VL tracking is employing customized and heavier unimodal encoders, and multi-modal fusion models. Albeit effective, existing VL trackers separate feature extraction and feature integration, resulting in extracted features that lack semantic guidance and have limited target-aware capability in complex scenarios, e.g., similar distractors and extreme illumination. In this work, inspired by the recent success of exploring foundation models with unified architecture for both natural language and computer vision tasks, we propose an All-in-One framework, which learns joint feature extraction and interaction by adopting a unified transformer backbone. Specifically, we mix raw vision and language signals to generate language-injected vision tokens, which we then concatenate before feeding into the unified backbone architecture. This approach achieves feature integration in a unified backbone, removing the need for carefully-designed fusion modules and resulting in a more effective and efficient VL tracking framework. To further improve the learning efficiency, we introduce a multi-modal alignment module based on cross-modal and intra-modal contrastive objectives, providing more reasonable representations for the unified All-in-One transformer backbone. Extensive experiments on five benchmarks, i.e., OTB99-L, TNL2K, LaSOT, LaSOTExt and WebUAV-3M, demonstrate the superiority of the proposed tracker against existing state-of-the-art (SOTA) methods on VL tracking. Codes will be available at https://github.com/983632847/All-in-One here.

From May 2019, GitHub launched sponsor mechanism indicating that GitHub is moving towards deeper integration of open source development and economic support. It will bring more comprehensive and diversified support to the open source community. However, the number of developers profiting from the sponsor mechanism follows a long tail distribution. Our study found that only 31% of developers who started the sponsor mechanism received rewards, and 39.3% of them only received a reward of one dollar. Our work focuses on identifying what factors affect the availability of sponsorship for developers in open source community. We start by defining 45 features to characterize the developers in four dimensions i.e. Personality, Advertisement, Repository and Behavior. The results of statistical analysis indicate that most of the proposed features differ significantly between the ones who received rewards (short for MTs_Yes) from those that are not. After that, we build machine learning model based on the proposed features to predict MTs_Yes. Compared with the existing work, results show that our method outperforms baselines by 30% for AUC (Area Under the Curve). In addition, we investigated the relative contribution of features in detecting MTs_Yes and analyzed the important features by using an interpretable model SHAP. Finally, based on the experimental results, we put forward corresponding and practical suggestions for developers who want to receive rewards so as to make the community of open source projects develop more harmonious.